Podcast App Redesign

Integrating machine learning and improved sharing features into the native iOS Podcast App for a more personalized experience.


Design Goals

  • Make discovering new episodes and podcasts easier through machine learning suggestions
  • Make sharing clips of episodes easier through natural language processing and crowd-sourced suggestions

Who benefits?

Podcast Listeners

Podcast listeners will have more personalized recommendations for new episodes and podcasts that they may not have discovered otherwise. This increases engagement by providing new sources of podcasts in an otherwise stagnant top charts list. The improved sharing experience makes it easy to share episodes, but now they can also share clips/snippets of episodes through the transcript highlights. This is made easier through collecting other user highlights to show top highlights that are likely to be shared so the listener doesn't need to scroll through the entire transcript to find what they are likely to want to share. 



Currently within the Podcast App, there is only featured and top chart podcasts, which makes it difficult for lesser known podcasts to gain an audience. Having discovery based off of listener behavior and not who has the most listeners gives indie podcasters more incentive to make their podcasts available on the app. For all podcasters, they will also have more data available to them so they can understand their listener's behaviors. By keeping tabs on drop rates, sharing, and subscriptions, podcasters can form personas on who listeners to their podcast and make informed decisions on how they can craft to their audiences. 


The App Engineers and Product Team

One of the purposes of machine learning is to take the human effort out of doing otherwise tedious and mundane tasks. Many other platforms like Facebook, Spotify, and Instagram already give you recommendations based on your natural behavior and actions, so the Podcast App can utilize similar functions to do the same. This involves little to no maintenance on the part of engineers once the algorithm and design is built out, and provides a beneficial feature to the audience, creators, and business. 


Apple Business Side

Podcasting and streaming is not isolated to iTunes anymore. There are streaming and hosting competitors in the market such as Spotify, SoundCloud, and Audible as well as podcast apps like Overcast and NPR One, which provide more features. If Apple ever wants to significantly monetize on hosting podcasts, they'll need to maintain their market share which means making Podcasters and Listeners happy. 


Interactive Data Model

The left side consists of the data points that are needed to create the right side personalization features that will be parts of the UI within the app. The data points are separated into 3 categories: individual user input, collective user input, and user behavior. 

Individual user input: An individual's action that requires conscious effort or thought. 

Collective user input: A conglomeration of actions from a multitude of users that require conscious effort or thought. 

User behavior: Actions that require little effort or no effort at all on the part of the user. These can be reactions to user input such as subscribing to a podcast creates a reaction of a podcast subscription becoming inactive over time.  Skipping episodes is a user action, but requires little effort on the part of the user.

Data Point Definitions:

  1. Likes/dislikes per episode*: The user generated rating accessible for each episode. Since rating an episode requires an extra step by the user to make a conscious decision, rating should be given more weight than unconscious behavior. 
  2. Podcasts subscribed to: The podcasts that a user is subscribed to, visible under "My Podcasts."
  3. Twitter hashtags associated with shares*: Words and sentiment associated with the hashtags when a user shares a podcast, episode, or clip to twitter or other social network sites with hashtags. These are used to determine sentiment and group together episodes and podcasts for auto-generated playlists. 
  4. Episodes shared: An episode that the listener selects the share button and publishes outside of the app. Since sharing episodes requires much more effort on the part of the user, these episodes should be given heavy weight in terms of the listener's favorability of the episode. 
  5. Transcript highlights*: The count of the highlights of a section of the transcript from an episode is combined up from all users. When a section is highlighted by multiple different people, then that section of the transcript will be shown as a "top highlight" or "suggested highlight."
  6. What other people who like what you like listen to*: Associating that people who liked that episode of This American Life also tended to like an episode of Freakonomics.
  7. Episodes skipped: The episodes that you never get around to listening to and the episodes you skip over when they come on. Skipping an episode is a way of saying that you don't want to listen to this now or repeatedly skipping a certain podcast's episodes may mean that your tastes are changing.
  8. Episode completion: The status of an episode being un-listened to, partially listened to, or completed at what point the listener stopped listening. 
  9. Active subscriptions vs. inactive subscriptions: After some time, if you too many un-listened to episodes that are building up in your queue for a particular podcast then your subscription is considered inactive. 
  10. What you listen to at certain times of day*: Perhaps you like to listen to news type podcasts in the early morning and fictional story podcasts like Welcome to Night Vale in the evening. This combination of a time stamp with what you are listening to will help create specific time related playlists.

*A new design implementation that does not use current existing features.


Machine Learning Features 


Rate episodes to help the algorithm learn your preferences 

The app will learn your preferences based on things you naturally do such as skipping episodes, episode completion rate, and inactive vs. active subscriptions, but it helps to include intended user input. The rate button provides a binary data input of a user having a positive or negative reaction to an episode. The algorithm will take this information and create playlists and suggestions based on your preferences.   


DiscoverY personalized just for yoU

Episodes Just For You: Similar to Spotify's Discover Weekly, the system will create a personalized playlist of episodes based on listening behavior and your ratings. It will also compare collective user data to recommend episodes that other people who like what you like are listening to. 

Podcasts Recommended For You: Featured and top charts perpetuates everyone to listen to the same podcasts, but with machine learning your personal taste will be taken into account. 

Playlists Recommended For You: The native iOS app doesn't currently support playlists, which is a sorely missed feature. For those who want to discover episodes related to certain topics or moods, the system will suggest playlists that it thinks you may be interested in. Playlists may also be auto-generated from the system based on transcript data and twitter hashtags associated with sharing an episode or clip. 

Your Commute Playlist: Sometimes we like to listen to certain types of episodes during certain times of the day. The commute playlist breaks down your listening patterns based on what you listen to during the morning or evening. 


Sharing clips is now possible with suggested top highlights  

Sharing an entire 45 minute episode with a friend may mean they never listen to it, but sharing a short clip will increase the odds that they'll take the time to check it out. This is great for podcasters who will have more exposure of their podcasts to new audiences. 

With natural language processing, transcripts can be created for each episode. To increase accuracy, podcasters can upload their own transcripts, which many of them already create per episode. This will help the system learn and refine it's natural language processing with huge sets of data. 

Similar to the way The Kindle or Medium shows top highlights from other users, the system will also show the top highlights to make it easier to share clips that you are likely to want to share. 


Take Aways

  • Personalized recommendations will increase listener engagement by keeping their feed fresh and will help podcasters gain exposure beyond the semi-stagnant "top charts" list.
  • Sharing is more powerful than ever with the ability to share clips instead of entire episodes that others are unlikely to listen to in its entirety.
  • Machine learning makes all of these features possible with little human maintenance and input.
Screen compilation.png